Robust Skin Lesion Segmentation Using a Hybrid CNN Model and Compound Scaling

Abstract

Medical image processing by computers has become an area of growing importance, with researchers shifting their attention to various groundbreaking state-of-the-art models that aim to ease the pressure on doctors. Notwithstanding, image segmentation and feature extraction are critical yet challenging tasks in clinical diagnostics, hindered by issues such as noise and artifacts, limited annotated data, and the need for high computational efficiency. This research focuses on developing an EfficientNet-U-Net model, which aims to improve medical image segmentation for detecting skin lesions utilizing the International Skin Imaging Collaboration Challenge (ISIC) 2018 dataset. The model combines the EfficientNet encoder, optimized for proficient feature extraction via compound scaling, with a U-Net-style decoder that uses transposed convolutions and skip connections to preserve spatial information. The approach also utilizes a hybrid loss function that combines Dice and Binary Cross-Entropy, as well as an adaptive learning rate schedule. Experimental results indicate that the proposed model performs well for the fully automated identification and delineation of skin lesions with the potential for clinical use. The proposed model achieves performance comparable to that of existing skin lesion segmentation methods.